Endmember Learning with K-Means through SCD Model in Hyperspectral Scene Reconstructions

Area of Science:

• Signal Processing
• Machine Learning
• Remote Sensing

Background:

• Compressive sensing (CS) enables sparse signal decomposition using dictionaries.
• Dictionary learning (DL) quality directly impacts CS reconstruction accuracy, especially in hyperspectral imaging (HSI).
• Existing methods like classic sparse coding dictionary (C-SCD) often rely on random sampling, limiting efficiency.

Purpose of the Study:

• To propose a novel and efficient dictionary learning method for CS applications.
• To enhance the performance of HSI scene reconstruction using improved dictionary learning.
• To develop a K-Means Sparse Coding Dictionary (KMSCD) for superior accuracy and speed.

Main Methods:

• Utilized K-Means clustering to derive dictionary centers from input data.
• Employed a greedy approach, integrating K-Means with orthogonal matching pursuit (OMP), for dictionary element learning.
• Evaluated KMSCD performance on publicly available HSI datasets for scene reconstruction.

Main Results:

• KMSCD demonstrated ~40% higher accuracy, 5x faster convergence, and double the robustness compared to C-SCD.
• Reconstructions using KMSCD showed 20-500% better mean accuracies than competing algorithms across five datasets.
• KMSCD improved trace material recovery by ~12% over C-SCD.
• Integration with Fast non-negative orthogonal matching pursuit (FNNOMP) yielded 10x better results than TMM for material allocation.

Conclusions:

• The proposed KMSCD significantly enhances HSI scene reconstruction quality and efficiency.
• KMSCD offers a robust and fast alternative to traditional dictionary learning methods.
• KMSCD combined with FNNOMP shows promise for material allocation in HSI simulators.